Detection of Ventricular Septal Defect in Pediatric Cardiac Ultrasound Videos Using Parasternal View and Faster R-CNN

Abstract

Congenital heart disease (CHD), particularly ventricular septal defect (VSD), remains a major contributor to pediatric morbidity, while echocardiographic diagnosis is highly dependent on operator expertise and image quality. This study examines the feasibility of an object-detection-based intelligent imaging framework for localizing VSD in pediatric cardiac ultrasound videos acquired from the parasternal long-axis view. Rather than proposing a novel detection algorithm, this work adopts a system-oriented approach by evaluating the Faster R-CNN framework under practical clinical constraints, including limited annotated data and heterogeneous ultrasound characteristics. Three convolutional neural network backbones such as ResNet50, ResNet101, and Inception-ResNet V2 are comparatively analyzed within a unified detection pipeline. Experimental results indicate that the ResNet101-based model achieves the highest localization performance at an intersection-over-union threshold of 0.5, while ResNet50 provides more consistent precision across stricter localization thresholds. Although false-positive detections are observed in acoustically challenging frames, the proposed framework maintains real-time feasibility at approximately 7–8 frames per second. The findings offer practical insights into accuracy–efficiency trade-offs and backbone selection for the development of clinically aware intelligent echocardiography systems, supporting the application of information and communication technology in pediatric cardiac imaging.